A method and system for fire prevention detection in a ship's hold

Abstract

The application provides a ship cabin fire prevention detection method and system, comprising the following steps: collecting atmospheric environment data in the ship cabin, obtaining data of vehicle internal sensors in the ship cabin, and calculating three-dimensional space coordinates of each vehicle internal sensor in a unified coordinate system of the ship cabin; taking each vehicle internal sensor as a graph node, constructing edges based on heat conduction relationships between the graph nodes, and generating a dynamic space-time graph G(t) capable of representing a thermodynamic system of the ship cabin; performing multi-level anomaly detection based on the dynamic space-time graph G(t), and performing fire risk assessment according to a detection result. The application performs overall modeling on a thermodynamic system of the whole ship cabin, performs prior detection by using vehicle internal sensor data, models and analyzes cluster risks under dense parking of vehicles, simultaneously outputs alarm signals of different levels by using hierarchical alarm logic, and can realize more efficient fire prevention detection.

Description

Technical Field

[0001] This invention belongs to the field of fire prevention system technology, specifically relating to a fire prevention detection method and system for ship cabins. Background Technology

[0002] Currently, automobiles transported on roll-on / roll-off ships mostly rely on passive response methods such as smoke detectors and video image recognition. These methods typically trigger alarms only after a large amount of smoke or open flame has already formed, resulting in significant response delays and a high risk of missing early warning windows. Furthermore, existing technologies lack the ability to predict the latency period of thermal runaway and struggle to model and analyze the cluster risk of fires spreading rapidly through thermal radiation and convection when vehicles are densely parked. With the surge in export demand for electric vehicles powered by lithium batteries, the probability of thermal runaway during transoceanic transport has greatly increased, significantly raising the fire risk, and existing protection systems are no longer sufficient to meet the demands. Summary of the Invention

[0003] This invention is made to solve the above-mentioned problems, and aims to provide a method and system for fire prevention and detection in ship cabins, which performs overall modeling of the thermodynamic system of the entire ship cabin to achieve more efficient fire prevention and detection.

[0004] This invention provides a method for fire prevention and detection inside a ship's cabin, characterized by the following steps: collecting atmospheric environmental data inside the cabin, acquiring data from vehicle interior sensors inside the cabin, and calculating the three-dimensional spatial coordinates of each vehicle interior sensor in a unified coordinate system within the cabin; using each vehicle interior sensor as a graph node, constructing edges based on the thermal conduction relationship between the graph nodes, and generating a dynamic spatiotemporal graph G(t) that characterizes the thermodynamic system of the cabin; performing multi-level anomaly detection based on the dynamic spatiotemporal graph G(t), and conducting a fire risk assessment based on the detection results.

[0005] In one embodiment of the present invention, in the step of constructing edges based on the heat conduction relationship between graph nodes, the edges include: in-vehicle edges, which establish in-vehicle edge connection relationships between graph nodes inside the same vehicle, and the in-vehicle edge connection relationships and in-vehicle edge weights are determined based on a preset topology model of in-vehicle sensors; and inter-vehicle edges, which establish inter-vehicle edge connection relationships between graph nodes in different vehicles, and the inter-vehicle edge weights of graph nodes are dynamically determined based on real-time spatial distance and atmospheric environment data between graph nodes.

[0006] In one embodiment of the present invention, the weight of the inner edge of the vehicle is a fixed value, and the calculation method is as follows: ,in, Indicates the weight of the edges inside the vehicle. Indicates thermal resistance.

[0007] In one embodiment of the present invention, the weight of the vehicle-to-vehicle edge is a dynamic value, and the calculation method is as follows: ,in, Let 'a' represent the inter-vehicle edge weight at time 't', 'a' represent the temperature influence coefficient, and 't' represent the time step. d represents the ambient temperature inside the ship's cabin. ij This represents the real-time spatial distance between graph nodes.

[0008] In one embodiment of the present invention, the dynamic spatiotemporal graph G(t) is constructed as follows: G(t) = (V, E(t), X(t)), where Represents the set of graph nodes. express The feature matrix of graph nodes at time step 1. express The set of edges at time points.

[0009] In one embodiment of the present invention, the steps of multi-level anomaly detection based on a dynamic spatiotemporal graph G(t) include: single-node anomaly detection, performing real-time statistical analysis on the features of each graph node to obtain feature values, and triggering a single-point anomaly event when the feature value exceeds a first preset threshold; local subgraph anomaly detection, performing clustering or subgraph analysis on local regions in the dynamic spatiotemporal graph G(t), and triggering a local subgraph anomaly event when anomaly node clusters are identified; and global graph anomaly detection, constructing a spatiotemporal graph autoencoder model, inputting the dynamic spatiotemporal graph G(t) into the spatiotemporal graph autoencoder model, calculating the reconstruction loss of the spatiotemporal graph autoencoder model, and triggering a global graph anomaly event when the reconstruction loss exceeds a second preset threshold.

[0010] In one embodiment of the present invention, the training process of the spatiotemporal graph autoencoder model includes: using historical spatiotemporal graph data from vehicle internal sensors as the training set, wherein the historical spatiotemporal graph data only includes data under normal operating conditions; setting a loss function for reconstruction loss; updating the internal parameters of the spatiotemporal graph autoencoder model through backpropagation algorithm and optimizer to minimize reconstruction loss.

[0011] In one embodiment of the present invention, alarm signals of different levels are output according to the triggering of single-point abnormal events, local subgraph abnormal events and global graph abnormal events, in accordance with a preset hierarchical alarm logic.

[0012] In one embodiment of the present invention, the hierarchical alarm logic includes: if only a single-point abnormal event is triggered, a level 3 alarm is output; if a single-point abnormal event and a local subgraph abnormal event are triggered simultaneously, but a global graph abnormal event is not triggered, a level 2 alarm is output; if a single-point abnormal event, a local subgraph abnormal event, and a global graph abnormal event are triggered simultaneously, a level 1 alarm is output.

[0013] This invention also provides a ship cabin fire prevention and detection system, characterized by the following features, derived from the aforementioned ship cabin fire prevention and detection method: a data acquisition module, which collects atmospheric environmental data within the ship cabin, acquires data from vehicle interior sensors within the ship cabin, and calculates the three-dimensional spatial coordinates of each vehicle interior sensor in a unified coordinate system within the ship cabin; a dynamic spatiotemporal graph construction module, which uses each vehicle interior sensor as a graph node, constructs edges based on the thermal conduction relationship between graph nodes, and generates a dynamic spatiotemporal graph G(t) that characterizes the thermodynamic system of the ship cabin; and a multi-level anomaly detection module, which performs multi-level anomaly detection based on the dynamic spatiotemporal graph G(t) and conducts fire risk assessment based on the detection results.

[0014] The role and effect of invention

[0015] According to the fire prevention and detection method and system for ship cabins involved in this invention, the invention constructs a dynamic spatiotemporal map by fusing vehicle interior sensor data and atmospheric environmental data, and performs anomaly detection at three levels: single-node anomaly detection, local subgraph anomaly detection, and global graph anomaly detection. Compared with existing technologies, this invention can utilize vehicle interior sensor data for pre-emptive detection, model and analyze the cluster risk of densely parked vehicles, and simultaneously employ hierarchical alarm logic to output alarm signals of different levels, effectively solving the problems of delayed fire detection, lack of predictive ability, and inability to analyze cluster risks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of the fire prevention and detection method inside the ship's cabin in an embodiment of the present invention.

[0018] Figure 2 This is a flowchart of the hierarchical alarm logic in an embodiment of the present invention.

[0019] Figure 3 This is a block diagram of a fire prevention and detection system inside a ship's cabin, as described in an embodiment of the present invention.

[0020] Explanation of reference numerals in the attached figures 100-Fire prevention and detection system inside the ship's cabin; 101-Data acquisition module; 102-Dynamic spatiotemporal map construction module; 103-Multi-level anomaly detection module; 104-Alarm logic module. Detailed Implementation

[0021] The technical solutions disclosed in this invention will be described in detail below with reference to specific embodiments.

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be noted that the illustrations provided in this embodiment are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] In this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used only for descriptive and distinguishing purposes and should not be construed as indicating or implying relative importance.

[0025] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the fire prevention and detection method and system for ship cabins of the present invention.

[0026] like Figures 1-3 As shown, this invention provides a method and system for fire prevention and detection in ship cabins based on dynamic spatiotemporal maps and multi-level anomaly detection. This method and system integrate vehicle interior sensor data and atmospheric environmental data, constructing a dynamic spatiotemporal map and performing single-node anomaly detection, local subgraph anomaly detection, and global map anomaly detection to achieve a shift from passive response to proactive prevention. This method and system for fire prevention and detection in ship cabins can be widely applied in various ship scenarios transporting electric vehicles, and is particularly suitable for car carriers, passenger ro-ro ships, and other situations requiring early warning of battery thermal runaway risks under high-density vehicle parking conditions.

[0027] like Figure 1 As shown, the fire prevention and detection method inside a ship's cabin provided by the present invention includes the following steps: S1, collect atmospheric environment data inside the cabin, acquire data from the sensors inside the vehicles inside the cabin, and calculate the three-dimensional spatial coordinates (x, y, y) of each vehicle's internal sensor in the unified coordinate system of the cabin. i y i , z i As a preferred option, given the high risk of battery thermal runaway in electric vehicles, in order to better model and predict the overall fire risk of the vehicle, the in-vehicle sensor can be the vehicle's battery management system sensor i.

[0028] In the specific implementation process, the first step is to acquire multi-source data and perform spatial coordinate mapping, collecting atmospheric environmental data within the ship's cabin. This data includes, but is not limited to, temperature, humidity, and carbon monoxide concentration. This atmospheric environmental data can be obtained in real time through various atmospheric environmental sensors deployed within the ship's cabin. Simultaneously, data from the battery management system sensors of each electric vehicle within the cabin is acquired. This data includes, but is not limited to, sensor identification, cell temperature, and voltage, and can be directly parsed from the battery management system data stream, including temperature T. i Temperature change rate Under ideal data conditions, battery management system (BMS) sensor data can be acquired in real time through the data interface between the vehicle and the ship, and transmitted in real time to the ship's data processing center. It is worth noting that this application does not limit the specific number or type of data interfaces. There can be one or more data interfaces, and different types of data interfaces can have the same or different functions, specifically ensuring stable transmission of BMS sensor data. For example, wireless communication can include, but is not limited to, Wi-Fi, Bluetooth, ZigBee, ultra-wideband, or cellular network communication. For example, it can include a wireless access point deployed inside the ship's cabin establishing a connection with the vehicle's built-in wireless communication module to achieve wireless transmission of BMS sensor data. Wired connections can include, but are not limited to, controller area network (MAN) bus connections, Ethernet connections, or universal serial bus connections. For example, it can include the vehicle establishing a physical connection with the ship's data acquisition terminal via a dedicated wiring harness when loaded, transmitting BMS data streams via MAN bus protocols or Ethernet protocols. This invention does not impose specific limitations on the communication protocols of the data interfaces, as long as reliable transmission of BMS sensor data can be achieved.

[0029] After acquiring the battery management system sensor data, it is necessary to calculate the three-dimensional spatial coordinates (x, y, y) of each battery management system sensor i in the unified coordinate system of the ship's cabin. i y i , z iThis facilitates subsequent graph structure construction. This invention does not impose specific restrictions on the calculation method of three-dimensional spatial coordinates, as long as the actual spatial position of each battery management system sensor i within the cabin can be accurately obtained. As one implementation method, a vision-based approach can be used to calculate the coordinates. Specifically, visible light cameras deployed inside the cabin capture images, and object detection algorithms are applied to process the captured images to identify the position and orientation of each vehicle in a unified coordinate system within the cabin. The object detection algorithm can be, for example, but not limited to, YOLO, Faster R-CNN, etc., as long as it can accurately identify the vehicle's position and orientation. Subsequently, from a pre-defined vehicle model database, the topology model of its battery management system sensor i is retrieved based on the vehicle's identity. This topology model defines the three-dimensional offset of each battery management system sensor i within the vehicle relative to a vehicle reference point. The vehicle reference point can be, for example, but not limited to, the vehicle's centroid or geometric center. Finally, combining the vehicle's global position and the relative offset of the battery management system sensor i, the final three-dimensional spatial coordinates (x, y, y) of each battery management system sensor i are calculated. i y i , z i ).

[0030] As another implementation, an IoT-based approach can be used to calculate the three-dimensional spatial coordinates. Specifically, each vehicle is equipped with an IoT positioning unit during loading. This IoT positioning unit can be, for example, but not limited to, an ultra-wideband positioning tag or an active RFID tag. The three-dimensional spatial coordinates of each IoT positioning unit are acquired in real time through positioning base stations deployed in the cabins. Since each IoT positioning unit corresponds one-to-one with a vehicle, the three-dimensional spatial coordinates of the vehicle are obtained. Subsequent steps are the same as the vision-based approach: the topology model of the vehicle's battery management system sensor i is retrieved from the vehicle model database, and the precise three-dimensional spatial coordinates (x, y, x) of each battery management system sensor i are calculated by combining the vehicle's global position and the relative offset of the sensor. i y i , z i ).

[0031] It is worth noting that this invention does not limit the specific methods of vehicle identification and sensor coordinate calculation. Vehicle identification can be obtained through various methods such as visual recognition, radio frequency identification, and manual input. Sensor coordinate calculation can also employ other spatial positioning technologies, specifically based on the ability to accurately obtain the three-dimensional spatial coordinates of each battery management system sensor i in the unified coordinate system of the ship's cabin. Furthermore, this invention does not limit the specific number and type of battery management system sensors i. The number of battery management system sensors i can vary between different vehicles. The types of sensors can include temperature sensors, voltage sensors, current sensors, etc., specifically based on their ability to reflect the thermal state of the battery cells.

[0032] S2 uses each battery management system sensor as a graph node, and constructs edges based on the thermal conduction relationships between the graph nodes to generate a dynamic spatiotemporal graph G(t) that characterizes the thermodynamic system of the ship's cabin. Here, a graph node refers to a node in a graph structure constructed using the battery management system sensors as basic units. A dynamic spatiotemporal graph is a graph structure that changes dynamically over time, where the node characteristics and edge weights can change with time.

[0033] This invention uses each battery management system sensor as a graph node, constructs edges based on the thermal conduction relationships between graph nodes, and generates a dynamic spatiotemporal graph G(t) that characterizes the ship's thermodynamic system. The construction method of the dynamic spatiotemporal graph G(t) is as follows: G(t) = (V, E(t), X(t)), where V represents the set of graph nodes, X(t) represents the graph node feature matrix at time t, E(t) represents the set of edges at time t, and the node features in X(t) may include temperature T and the rate of temperature change. The aforementioned features can be directly parsed from the battery management system sensor data stream. In the step of constructing edges based on the thermal conduction relationships between graph nodes, edges include two types: in-vehicle edges and inter-vehicle edges. In-vehicle edges refer to the connections established between graph nodes within the same vehicle, used to characterize the physical thermal conduction characteristics between battery cells within the vehicle. The connections of in-vehicle edges are determined based on the topology model of battery management system sensor i retrieved from the vehicle model database. The topology model of battery management system sensor i defines the physical connections between various sensor nodes within the same vehicle. As one implementation, sparse graph in-vehicle edges are established only between physically directly adjacent sensor nodes, such as connections between sensor nodes corresponding to adjacent battery cells. This connection method accurately reflects the thermal conduction path between battery cells. As another implementation, fully connected graph in-vehicle edges can be established between all sensor nodes in the same vehicle. This connection method involves a larger computational load but can more comprehensively reflect the thermal conduction relationships within the vehicle. The weight of the in-vehicle edges uses a preset fixed value W. A This characterizes the physical heat conduction properties within the nodes. A fixed value W... AThe calculation method can be based on a physical model, such as W. A =1 / R(th), where R(th) represents the thermal resistance, which can be calculated based on parameters such as the material properties of the battery cell, contact area, and contact pressure. This fixed value W A It can be pre-calculated and stored in a vehicle model database, and retrieved directly when needed. As another implementation, this fixed value W... A It can also be an empirical constant obtained from thermodynamic simulation or experimental measurement; this invention does not limit the fixed value W. A The specific calculation method only needs to characterize the physical heat conduction characteristics between nodes. Vehicle edges refer to the connections established between graph nodes belonging to different vehicles, used to characterize the heat exchange relationships between different vehicles through thermal radiation and convection. The establishment of vehicle edges requires satisfying the spatial proximity condition, that is, only when the three-dimensional spatial distance is less than a certain preset threshold. Establish connections between graph nodes belonging to different vehicles. This preset threshold... The threshold value can be, for example, but is not limited to, 1 meter. This invention does not limit the threshold value. The specific values ​​can be adjusted based on the vehicle parking density and heat conduction characteristics within the ship's hold. The weight of the vehicle-to-vehicle edge uses a dynamic weight W. B The dynamic weight W is used to characterize the efficiency of heat radiation and convection exchange between vehicles. B The global atmospheric environment parameter is dynamically determined at each time step t based on the real-time spatial distance between graph nodes and at least one global atmospheric environment parameter. This global atmospheric environment parameter may include, but is not limited to, the ambient temperature T measured by the cabin atmospheric sensor. amb Airflow velocity state V air etc. As one implementation method, the inter-vehicle edge weight W B The calculation of (t) uses linear or nonlinear functions, for example Where a represents the temperature effect coefficient, C is a constant, and d ij This represents the real-time spatial distance between graph nodes. The function indicates that the higher the ambient temperature and the stronger the thermal radiation, the greater the weight of the vehicle-to-vehicle edge; the closer the spatial distance, the higher the heat exchange efficiency, and the greater the weight of the vehicle-to-vehicle edge. As another implementation, the vehicle-to-vehicle edge weight W... B The calculation of (t) can be achieved by interpolation using a pre-defined lookup table. This lookup table stores weighting coefficients for different combinations of environmental parameters and can be generated through offline computational fluid dynamics (CFD) simulation. The advantage of this method is that it can more accurately reflect the heat exchange relationship under complex environmental conditions.

[0034] It is worth noting that this invention does not limit the specific form of the dynamic function for the inter-vehicle edge weights, as long as the weights of the inter-vehicle edges can be dynamically determined based on the spatial distance between nodes and real-time environmental parameters to characterize the relationship between external thermal radiation and convection. Similarly, this invention does not limit the specific numerical determination method for the weights of the in-vehicle edges, with a fixed value W. A It can be a theoretical value calculated based on a physical model, or an empirical value based on experimental measurements, depending on whether it can characterize the internal physical heat conduction characteristics between nodes.

[0035] S3 performs multi-level anomaly detection based on a dynamic spatiotemporal graph G(t), and conducts fire risk assessment based on the detection results. The multi-level anomaly detection includes three levels: single-node anomaly detection, local subgraph anomaly detection, and global graph anomaly detection. These three levels assess fire risk from three different scales, complementing each other to form a complete fire prevention and detection system.

[0036] Single-node anomaly detection refers to real-time statistical analysis of the features of each graph node to obtain feature values. When a feature value exceeds a first preset threshold, the anomaly is detected. and When this occurs, a single-point anomaly event is triggered. The characteristics of the graph nodes include temperature T and the rate of temperature change. The characteristics of these graph nodes can directly reflect the thermal state of individual battery cells. Real-time monitoring of the temperature change rate of all graph nodes is also possible. And temperature T, if the rate of temperature change Greater than Or temperature T greater than If so, it is considered that the current single node has triggered an abnormal event. First preset threshold. and It can be set based on historical data and experience values, for example It can be set to 2℃ / min. It can be set to 60℃, and the specific value can be adjusted according to the battery type and environmental conditions. Single-node anomaly detection is a high-frequency detection method that can quickly capture abnormal temperature rises in individual cells, which is the basis for early warning.

[0037] Local subgraph anomaly detection refers to performing cloud clustering or subgraph analysis on local regions within a dynamic spatiotemporal graph G(t). When anomaly node clusters are identified, a local subgraph anomaly event is triggered. The purpose of local subgraph anomaly detection is to identify anomalous sensor clusters within a small area. When multiple neighboring sensor nodes simultaneously exhibit anomalies, this often indicates that a fire risk is forming in the local area. This invention does not limit the specific algorithm for local subgraph anomaly detection. As one implementation, a point cloud clustering algorithm is used to perform clustering analysis on the multidimensional features of all nodes. If a high-risk cluster with a density exceeding a threshold is found, a local subgraph anomaly event is considered to exist. The point cloud clustering algorithm can be, for example, but not limited to, the DBSCAN algorithm, and the multidimensional features can be... Temperature T, rate of temperature change The DBSCAN algorithm can divide graph nodes into different clusters based on spatial distance and feature similarity. When the number of graph nodes in a cluster exceeds a preset threshold and the average temperature or average temperature change rate of the graph nodes in the cluster exceeds a preset threshold, the cluster is considered a high-risk cluster, triggering a local subgraph anomaly event. As another implementation, a graph neural network (GNN) subgraph analysis model can be applied to score anomalies in the subgraph composed of the nodes of interest and their K-order neighbors. If the anomaly score exceeds a preset threshold, a local subgraph anomaly event is considered triggered. The graph neural network subgraph analysis model can learn the topology and node characteristics of the subgraph, thereby more accurately determining whether there are anomalies in local areas. Local subgraph anomaly detection is a mid-frequency detection method, with a detection frequency lower than that of single-node anomaly detection. It is used to confirm and expand the results of single-point anomaly detection to determine whether there is cluster risk.

[0038] Global graph anomaly detection refers to the construction of a spatiotemporal graph autoencoder model, including model construction, offline model training, and online monitoring. In the offline model training phase, the dynamic spatiotemporal graph G(t) is input into the spatiotemporal graph autoencoder model, and the reconstruction loss of the model is calculated. When the reconstruction loss exceeds a second preset threshold, a global graph anomaly event is triggered. The spatiotemporal graph autoencoder model is a deep learning model based on graph neural networks, capable of being trained offline using a large amount of data under normal operating conditions to learn the normal thermodynamic linkage mode of the entire ship's cabin. In the online monitoring phase, the spatiotemporal graph autoencoder model reconstructs the model based on the currently input dynamic spatiotemporal graph. If the reconstruction error is large, it indicates a significant difference between the current thermodynamic linkage mode and the normal mode, i.e., a global anomaly exists.

[0039] Specifically, in the model construction phase, the spatiotemporal graph autoencoder model comprises two parts: an encoder and a decoder. The encoder consists of multiple stacked spatiotemporal graph convolutional layers, which can be, for example, but not limited to, a combination of graph convolutional networks and gated recurrent units (GCN+GRU), or a graph convolutional long short-term memory network (GCLSTM). The encoder compresses the input spatiotemporal sequence graph G(tk...t) into a low-dimensional latent vector z(t). The spatiotemporal sequence graph G(tk...t) represents a dynamic spatiotemporal graph sequence of k+1 consecutive time steps from time tk to time t, reflecting the temporal evolution of the thermodynamic system. The decoder consists of inverse graph convolutional layers, attempting to reconstruct the original graph sequence from the latent vector z(t). (tk...t). The decoder's structure is symmetrical to the encoder's, and it gradually restores the low-dimensional latent vectors to the original graph structure through deconvolution or unpooling operations.

[0040] During the offline model training phase, the offline training process of the spatiotemporal graph autoencoder model is as follows: First, historical spatiotemporal graph data of the battery management system sensors, containing only normal operating conditions, is used as the training set. Normal operating conditions refer to operating conditions under different environmental conditions, but without any fire incidents, including normal thermodynamic data under different ambient temperatures, airflow velocities, and vehicle parking densities. This data should cover as many normal operating condition scenarios as possible to ensure that the spatiotemporal graph autoencoder model can learn comprehensive normal linkage patterns. Second, the loss function L is set as the reconstruction loss, for example, the mean square error of node features. Where N is the total number of nodes, T i The original node temperature characteristics, The reconstructed node temperature features are used. The loss function L can also include the mean square error of the temperature change rate or the reconstruction error of other features, which can be set according to actual needs. Then, the internal parameters W of the spatiotemporal graph autoencoder model are updated through the backpropagation algorithm and optimizer. model The goal is to minimize the reconstruction loss L. The optimizer can be, for example, but not limited to, an adaptive moment estimator (Adam optimizer). After training, the spatiotemporal graph autoencoder model can be used to determine whether a time series graph is in a normal mode, and its internal parameters W... model It has been frozen and will no longer be updated.

[0041] During the online monitoring phase, the current dynamic spatiotemporal graph G is constructed in real time. live (t), the dynamic spacetime graph G live (t) includes the inter-vehicle edge weight W dynamically calculated in step S2. B (t). G live (t) Input the trained spatiotemporal graph autoencoder model for inference and prediction to obtain the reconstructed graph. live (t). Calculate the current reconstruction loss L in real time. live (t), if L live If (t) exceeds a preset second threshold, a global graph anomaly event is considered to exist. The second preset threshold can be determined based on the reconstruction loss distribution on the training set. For example, the second preset threshold can be set to three times the standard deviation of the average reconstruction loss of the training set. Global graph anomaly detection is a low-frequency detection method, and its detection frequency can be lower than that of single-node anomaly detection and local subgraph anomaly detection. It is used to assess whether there are abnormal patterns in the thermodynamic system of the entire ship compartment from a macroscopic level.

[0042] It is worth noting that this invention does not limit the specific network structure of the spatiotemporal graph autoencoder model, as long as it can achieve the encoding from dynamic spatiotemporal graph sequences to low-dimensional latent vectors and the reconstruction from latent vectors to graph sequences. Furthermore, this invention does not limit the specific form of the loss function; it can be mean squared error, cross-entropy loss, mean absolute error, etc.

[0043] S4 outputs alarm signals of different levels based on fire risk assessment. Specifically, based on the triggering of single-point anomalies, local sub-graph anomalies, and global graph anomalies, it outputs alarm signals of different levels according to a preset hierarchical alarm logic. This hierarchical alarm logic organically integrates the anomaly detection results of the three levels to form a clear and actionable risk assessment conclusion. For example... Figure 2 As shown, the hierarchical alarm logic is as follows: S4-1: If the concentration of a specified gas detected by real-time atmospheric monitoring exceeds the threshold, a Level 1 alarm will be triggered directly. A Level 1 alarm is the highest level alarm, indicating a significant anomaly in the thermodynamic system of the entire ship's compartment and a major fire risk. In this situation, operators should immediately take emergency measures, including activating the fire suppression system and organizing personnel evacuation. Specified gases may include carbon monoxide, hydrogen, and other gases released in the early stages of battery thermal runaway. A sudden increase in the concentration of these gases often indicates that a fire has already occurred or is about to occur; therefore, the highest level alarm should be triggered directly.

[0044] S4-2, if the concentration of a specified gas obtained from real-time atmospheric detection does not exceed the threshold: if only a single point abnormal event is triggered, a level 3 alarm will be output. The level 3 alarm is the lowest level alarm, used to remind the operator that one or several sensor nodes have abnormalities, but have not yet formed a local cluster risk. At this time, the operator can observe and pay attention, and does not take emergency measures for the time being.

[0045] S4-3, if the specified gas concentration obtained from real-time atmospheric detection does not exceed the threshold: if a single-point abnormal event and a local subgraph abnormal event are triggered simultaneously, but a global graph abnormal event is not triggered, a level 2 alarm will be output. The level 2 alarm is a medium-level alarm, indicating that there is an early fire risk in a local area, and multiple neighboring nodes are abnormal at the same time, but the overall thermodynamic system is still within the normal mode range. At this time, the operator should be more vigilant, strengthen the monitoring of the local area, and prepare to take corresponding preventive measures.

[0046] S4-4, If the concentration of a specified gas obtained from real-time atmospheric detection does not exceed the threshold: if a single-point abnormal event, a local subgraph abnormal event, and a global graph abnormal event are triggered simultaneously, a level one alarm will be output.

[0047] The hierarchical alarm process logic is designed hierarchically, with the alarm process proceeding from micro to macro. High-frequency, medium-frequency, and low-frequency methods are used to perform single-point anomaly detection, subgraph anomaly detection, and global anomaly detection on the dynamic spatiotemporal graph G(t). This hierarchical alarm logic combines redundancy and engineering feasibility, avoiding frequent alarms caused by single-point false alarms while issuing accurate warnings after multi-level confirmation, thus improving the system's reliability and practicality.

[0048] like Figure 3 As shown, the present invention also provides a ship cabin fire prevention and detection system 100, which is used to implement the above-mentioned ship cabin fire prevention and detection method. The system includes a data acquisition module 101, a dynamic spatiotemporal map construction module 102, a multi-level anomaly detection module 103, and an alarm logic module 104.

[0049] like Figure 3As shown, the data acquisition module 101 is used to collect atmospheric environmental data inside the cabin, acquire battery management system sensor data of vehicles inside the cabin, and calculate the three-dimensional spatial coordinates of each battery management system sensor in a unified coordinate system of the cabin. The data acquisition module 101 may include various sub-modules, such as an image acquisition sub-module, a target detection sub-module, a vehicle model library sub-module, and a coordinate calculation sub-module. Specifically, the image acquisition sub-module is used to capture images using visible light cameras deployed in the cabin; the target detection sub-module is used to process the images using target detection algorithms to identify the position and attitude of each vehicle in the unified coordinate system of the cabin; the vehicle model library sub-module is used to store the topology model of the battery management system sensor i for each vehicle model, which defines the three-dimensional offset of the battery management system sensor inside the vehicle relative to the vehicle's reference point; the coordinate calculation sub-module is used to combine the vehicle's global position and the relative offset of the sensor to calculate the final three-dimensional spatial coordinates of each battery management system sensor. As another implementation, the data acquisition module 101 may include an IoT positioning sub-module, used to acquire the three-dimensional spatial coordinates of the vehicle through an IoT positioning unit and a positioning base station.

[0050] like Figure 3 As shown, the dynamic spatiotemporal graph construction module 102 is used to construct edges based on the thermal conduction relationship between each battery management system sensor as a graph node, generating a dynamic spatiotemporal graph G(t) that can characterize the thermodynamic system of the ship's cabin. The dynamic spatiotemporal graph construction module 102 may include a graph node construction submodule, an in-vehicle edge construction submodule, an inter-vehicle edge construction submodule, and a weight calculation submodule. The graph node construction submodule is used to establish a set V of graph nodes, using each battery management system sensor as a graph node, and extract the node feature matrix X(t) of each node at time t. The in-vehicle edge construction submodule is used to establish in-vehicle edge connections between graph nodes within the same vehicle based on the topology model of the battery management system sensor i retrieved from the vehicle model database, and uses a preset fixed value W. A As the weight of the in-vehicle edge. The inter-vehicle edge construction submodule is used to construct edges in 3D space where the distance is less than a preset threshold. Establish inter-vehicle edge connections between graph nodes belonging to different vehicles, and use dynamic values ​​W. B As the inter-vehicle edge weight. The weight calculation submodule is used to dynamically determine the inter-vehicle edge weight W at each time step t based on the real-time spatial distance between graph nodes and global atmospheric environment parameters. B (t).

[0051] like Figure 3As shown, the multi-level anomaly detection module 103 is used for multi-level anomaly detection based on the dynamic spatiotemporal graph G(t), and performs fire risk assessment based on the detection results. Specifically, the multi-level anomaly detection module 103 may include a single-node anomaly detection submodule, a local subgraph anomaly detection submodule, a global graph anomaly detection submodule, and an alarm logic submodule. The single-node anomaly detection submodule is used to perform real-time statistical analysis of the characteristics of each graph node to obtain feature values. When the feature value exceeds a first preset threshold... and When an abnormal node cluster is identified, a single-point anomaly event is triggered. The local subgraph anomaly detection submodule is used to perform cloud clustering or subgraph analysis on local regions in the dynamic spatiotemporal graph G(t). When an abnormal node cluster is identified, a local subgraph anomaly event is triggered. The global graph anomaly detection submodule is used to construct a spatiotemporal graph autoencoder model. The dynamic spatiotemporal graph G(t) is input into the spatiotemporal graph autoencoder model, and the reconstruction loss of the spatiotemporal graph autoencoder model is calculated. When the reconstruction loss exceeds a second preset threshold, a global graph anomaly event is triggered.

[0052] like Figure 3 As shown, the alarm logic module 104 is used to output alarm signals of different levels according to the triggering conditions of the specified gas concentration, single-point abnormal event, local subgraph abnormal event and global graph abnormal event obtained by atmospheric detection, and according to the preset hierarchical alarm logic.

[0053] It is worth noting that the module division in this invention is only for descriptive convenience. In actual implementation, the modules can be integrated or further subdivided into more sub-modules. This invention does not limit the specific way the modules are divided, as long as the corresponding functions can be achieved. Furthermore, the steps in the various embodiments described above can be implemented using software, hardware, or a combination of both. Correspondingly, the modules in the above system can also be implemented using software, hardware, or a combination of both.

[0054] The present invention can also provide an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for fire prevention and detection inside the ship's cabin. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. The processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0055] This invention proposes a method and system for fire prevention and detection within ship cabins. By fusing sensor data from the battery management system and atmospheric environmental data, a dynamic spatiotemporal map is constructed. Anomaly detection is performed at three levels: single-node anomaly detection, local subgraph anomaly detection, and global graph anomaly detection. This achieves a shift from passive response to proactive prevention, effectively addressing the problems of poor fire detection timeliness, lack of predictive capabilities, and inability to analyze cluster risks in existing technologies. The dynamic spatiotemporal map constructed in this invention fully considers both in-vehicle and inter-vehicle heat conduction relationships, employing a combination of static and dynamic weights to accurately characterize the complex thermodynamic system within ship cabins. This invention uses a spatiotemporal map autoencoder model for unsupervised learning, enabling automatic learning of thermodynamic linkage patterns under normal operating conditions, and exhibiting high sensitivity to early-stage fire scenarios. The hierarchical alarm logic designed in this invention organically integrates the detection results from the three levels, avoiding interference from false alarms at a single level while providing accurate early warnings when risks actually materialize, demonstrating good robustness and engineering feasibility.

[0056] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0057] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for fire prevention and detection inside a ship's cabin, characterized in that, Includes the following steps: Collect atmospheric environmental data inside the cabin, obtain data from the vehicle's internal sensors inside the cabin, and calculate the three-dimensional spatial coordinates of each vehicle's internal sensor in the unified coordinate system of the cabin. Using each of the vehicle's internal sensors as a graph node, edges are constructed based on the thermal conduction relationships between the graph nodes to generate a dynamic spatiotemporal graph G(t) that can characterize the thermodynamic system of the cabin. Multi-level anomaly detection is performed based on the dynamic spatiotemporal graph G(t), and fire risk assessment is conducted based on the detection results.

2. The method for preventing and detecting fires inside a ship's cabin according to claim 1, characterized in that: In the step of constructing edges based on the heat conduction relationships between the graph nodes, the edges include: In-vehicle edges: In-vehicle edge connections are established between the graph nodes within the same vehicle. The in-vehicle edge connections and weights are determined based on a preset in-vehicle sensor topology model. Vehicle-to-vehicle edges are established between graph nodes of different vehicles. The weight of the vehicle-to-vehicle edge of the graph node is dynamically determined based on the real-time spatial distance between the graph nodes and the atmospheric environment data.

3. The method for preventing and detecting fires inside a ship's cabin according to claim 2, characterized in that: The weight of the inner edge of the vehicle is a fixed value, and the calculation method is as follows: ， in, Indicates the weight of the edges inside the vehicle. Indicates thermal resistance.

4. The method for preventing and detecting fires inside a ship's cabin according to claim 2, characterized in that: The weight of the vehicle-to-vehicle edge is a dynamic value, calculated as follows: ， in, This represents the inter-vehicle edge weight at time t, where 'a' represents the temperature influence coefficient and 't' represents the time step. d represents the ambient temperature inside the ship's cabin. ij This represents the real-time spatial distance between the graph nodes.

5. The method for preventing and detecting fires inside a ship's cabin according to claim 1, characterized in that, The method for constructing the dynamic spatiotemporal graph G(t) is as follows: G(t) = (V, E(t), X(t)), in Represents the set of graph nodes. express The feature matrix of graph nodes at time step 1. express The set of edges at time points.

6. The method for preventing and detecting fires inside a ship's cabin according to claim 1, characterized in that: The steps for multi-level anomaly detection based on the dynamic spatiotemporal graph G(t) include: Single-node anomaly detection involves performing real-time statistical analysis on the features of each graph node to obtain feature values. When the feature value exceeds a first preset threshold, a single-node anomaly event is triggered. Local subgraph anomaly detection involves clustering or subgraph analysis of local regions in the dynamic spatiotemporal graph G(t). When an abnormal node cluster is identified, a local subgraph anomaly event is triggered. Global graph anomaly detection involves constructing a spatiotemporal graph autoencoder model, inputting the dynamic spatiotemporal graph G(t) into the spatiotemporal graph autoencoder model, calculating the reconstruction loss of the spatiotemporal graph autoencoder model, and triggering a global graph anomaly event when the reconstruction loss exceeds a second preset threshold.

7. The method for preventing and detecting fires inside a ship's cabin according to claim 6, characterized in that, The training process of the spatiotemporal graph autoencoder model includes: Historical spatiotemporal map data from in-vehicle sensors is used as the training set, and the historical spatiotemporal map data only includes data under normal operating conditions. A loss function for reconstruction loss is defined, and the internal parameters of the spatiotemporal graph autoencoder model are updated through backpropagation algorithm and optimizer to minimize the reconstruction loss.

8. The method for preventing and detecting fires inside a ship's cabin according to claim 6, characterized in that, Also includes: Based on the triggering of the single-point anomaly event, the local subgraph anomaly event, and the global graph anomaly event, alarm signals of different levels are output according to a preset hierarchical alarm logic.

9. The method for preventing and detecting fires inside a ship's cabin according to claim 8, characterized in that, The hierarchical alarm logic includes: If only the single point of failure event is triggered, a level 3 alarm will be output; If the single-point anomaly event and the local subgraph anomaly event are triggered simultaneously, but the global graph anomaly event is not triggered, a level 2 alarm is output. If the single-point anomaly event, the local subgraph anomaly event, and the global graph anomaly event are triggered simultaneously, a level one alarm will be output.

10. A system for fire prevention and detection methods inside ship cabins according to any one of claims 1 to 9, characterized in that, include: The data acquisition module collects atmospheric environmental data inside the cabin, acquires data from the vehicle interior sensors inside the cabin, and calculates the three-dimensional spatial coordinates of each vehicle interior sensor in the unified coordinate system of the cabin. The dynamic spatiotemporal graph construction module uses each of the vehicle's internal sensors as a graph node, constructs edges based on the thermal conduction relationship between the graph nodes, and generates a dynamic spatiotemporal graph G(t) that can characterize the thermodynamic system of the cabin. The multi-level anomaly detection module performs multi-level anomaly detection based on the dynamic spatiotemporal graph G(t) and conducts fire risk assessment based on the detection results.

Images